Pad sequence by tokenizer in each batch

[kirilltobola]

For now we tokenize dataset before training.

We could tokenize and pad sequences by batch, using just tokenizer.